Blueprint Forge

Why it exists

After the US Government blocked users from using Fable 5 for almost three weeks, it showed that you can't assume that the high-tier model you have today will be there a week from now when you want to actually build the thing you have planned. This suite helps alleviate that problem by allowing you to batch-create blueprints once that are descriptive enough for cheaper models to follow months from now to create a result that is just about as good as if the high-tier model wrote it. The blueprinting process also accounts for the run order so that every blueprint is building off of what will exist when it is reached rather than simply what exists when it is created.

The interview is the other half. Every autonomous run eventually hits an ambiguous spec and resolves it with a coin flip, and in my blind benchmarks across 20+ graded runs, the hard interview stop caught every planted ambiguity in 8/8 runs while plain sessions asked zero questions and advisory gates self-waived under autonomous dispatch in 5/5.

There are, however, two caveats that are important to understand before using this. First, making any pivot in the project and its end state will likely invalidate any unused blueprints, so it is important that before blueprinting, you decide on a direction that you are 100% certain of, or you only blueprint up to a point that you are 100% certain of, else you risk wasting premium tokens on blueprints that may be invalidated. Second, this process is costlier and does not always ensure equal quality when compared to just using the high-tier model, though it can get close. That cost can be spread out across multiple sessions and even multiple subscriptions; I personally have the builder agents set to be GPT 5.6 models that run through headless Codex sessions.

Quick start

# in Claude Code — the repo is the plugin and its own marketplace
/plugin marketplace add SeanL128/blueprint-forge
/plugin install blueprint-forge@blueprint-forge

The first run bootstraps a blueprints/ folder in your repo with a MODELS.md routing map and a RUN-ORDER.md ledger. Three skills, all user-invoked only, because they deliberately spend premium planning-model tokens, so Claude never triggers them on its own.

• blueprint-oneshot One task end to end: interview, blueprint, your approval, then an autonomous build and review. Two hard stops, then hands-off. oneshot
• blueprint-backlog Turns a backlog or TODO into blueprints, each approved, revised, or rejected by you, plus a run-order ledger where every item assumes the state the previous ones leave behind. split · 1
• blueprint-implement Builds one approved blueprint: dispatches its builder subagents, reviews the result against it, then routes ship / fix here / re-dispatch / re-blueprint. split · 2

Configuration

Blueprints name roles, never model IDs, and blueprints/MODELS.md maps each role to a model for the harness in play. The default is an all-Claude split, and an optional per-role dispatch command column can send builds to another harness entirely. The split I run daily keeps blueprinting and review on Claude while GPT-5.6 builders run through headless Codex sessions, since building burns most of the tokens and a separate $20 ChatGPT subscription absorbs them; the worked map lives in the repo's usage doc.

• blueprintergrounds, interviews, writes and approves blueprints; the judgment layerpremium
• implementerruns the implement session: dispatches builders, runs verify commands, keeps the ledgermid
• builder-light / standard / heavyexecute blueprints cold, at the cheapest tier that can build the item wellcheap → mid
• reviewer / reviewer-heavychecks the diff against the blueprint's definition of done, escalating on riskindependent

How it works

A blueprint is a self-contained markdown file written for a builder that starts cold, works alone, and cannot ask questions, so if a fact isn't in the file or the repo, it doesn't exist for the builder. Each carries its builder role, grounding references, dependencies, starting state, a definition-of-done checklist, and reviewer notes. The ledger tracks a closed status vocabulary, and an item is only checked off once review passes, which makes the whole plan resumable across sessions, days, or models.

Every skill was built test-first against blind baseline runs, watching what context-free agents do wrong without the skill and then gating exactly that, and the suite was benchmarked across 20+ graded runs on toy and ~350-file repos before release.

Status & roadmap

Blueprint Forge is released and is the process I use daily for my own projects, but there are more updates that are planned for the future.

• Oneshot and split paths, benchmarked blind before releaseDone
• Worked Claude + GPT split via per-role Codex dispatch commandsDone
• Plugin packaging, with the repo as its own marketplaceDone

What I learned / what broke

Advisory process gates don't survive autonomy. Every suite I tested that recommends asking questions waived its own gate the moment it ran headless, five out of five, and only a hard stop-and-wait held. Building that took writing skills test-first against blind baseline runs rather than writing rules and hoping.

Blind graders penalize negotiated answers. The resolution the user actually chose in an interview can grade worse than the guess a judge would have made, so the suite's value shows up in variance, not means, which changed how I position it: insurance, not a quality upgrade.

The economics only work with routing. Pinned to one premium model, the pipeline costs 3 to 4 times a plain session, and arbitraging build tokens to cheaper builders cut the premium to about 1.5 times.